Qwen3.8-27B-MixQ5A6E8-v0 (experimental rc1)

Experimental candidate — not optimized, recommended, or generally validated. This immutable candidate exists to make a measured mixed-quant experiment reproducible. No quality-equivalence, speed-superiority, memory, or maximum-context claim is made beyond the sealed evidence cited below.

This community release is not affiliated with or endorsed by Qwen or Alibaba Cloud.

Artifact

File	Bytes	SHA256
`Qwen3.8-27B-MixQ5A6E8-v0.gguf`	20,804,373,120	`eae1e5ba33e6b3ebeda4a22dc540ae644f8114eb85df63848d80842cf78dd348`

Verify the download with sha256sum -c SHA256SUMS. MANIFEST.json is the canonical machine-readable provenance record.

Recipe and immutable provenance

Upstream model: Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 (Apache-2.0).
BF16 input: ggml-org/Qwen3.8-27B-GGUF@0669b98607d47046c7c2b3f801011d54a08cfccf, SHA256 5a3eedc837bcbd1365cdbf5b71e698df3122e76586ba872f07ce3ed4a9bfa97e.
Quantizer: llama.cpp b10442, commit 9b0a2ce859d3884252705e9ac7c93c98616bb238.
External imatrix: Bartowski Qwen3.8-27B-imatrix.gguf@f0eec4a4bb4975114a030d048952d83c0a53c034, SHA256 aaa933d4b9ce23e1f65c548ad34f16956d8af44a51b5c15bf4f393ba59508cd8.
Quantization: Q5_K_M fallback; Q6_K for matched .*attn.* and .*ssm.* tensors; Q8_0 token embedding and output. Tensor rules SHA256 cc2d2d8f797b3d6fb7895e0ec46a0dc1e88b5b9623d37d501385237183ae1033.

The external importance matrix was not created by this project. See ATTRIBUTION.md. Exact normalized dry-run/build argv and pins are retained in MANIFEST.json; the reproducible entry points are:

bash scripts/build_candidate.sh --mode dry-run
bash scripts/build_candidate.sh --mode build

Sealed Q5_K_M/Q6_K comparison (Vulkan lane)

Within this sealed three-artifact comparison only, MixQ5A6E8-v0 is a measured intermediate, nondominated point across artifact size, BF16-relative KL, and the three named llama-bench workloads. Lower is better for size and KL; higher is better for throughput.

Artifact	Bytes	BF16-relative KL	`pp512` median tok/s	`pp2048` median tok/s	`tg128` median tok/s
Q5_K_M	19,231,099,520	0.003949	1,141.60	1,668.81	39.0804
MixQ5A6E8-v0	20,804,373,120	0.003083	1,139.07	1,707.78	36.5086
Q6_K	22,082,529,920	0.002099	1,116.58	1,600.91	33.5946

Compared with Q6_K, the candidate is 5.7881% smaller and its medians are +2.0142% (pp512), +6.6756% (pp2048), and +8.6740% (tg128), while fidelity is worse: full PPL is 6.6674 versus 6.6553 and BF16-relative KL is 0.003083 versus 0.002099. Compared with Q5_K_M, it is 8.1809% larger, has lower BF16-relative KL (0.003083 versus 0.003949), and has mixed throughput: -0.2216% (pp512), +2.3352% (pp2048), and -6.5808% (tg128). Its full PPL of 6.6674 versus 6.6715 is too close to support a PPL-superiority claim.

Evidence for every value and the scoped classification above: results/llama-bench-mixq5a6e8-v0-vulkan-b10442-20260815T2111Z and results/ppl-full-bf16ngl48-mixq5a6e8-v0-vulkan-b10442-20260815T2210Z.

CUDA lane (sm_86, dual RTX 3090)

The identical three-model matrix was rerun with the same pinned b10442 tree compiled for CUDA sm_86 and the same protocol (pp512/pp2048/tg128, 5 repetitions, layer split 1/1). Median tok/s:

Artifact	`pp512`	`pp2048`	`tg128`
Q5_K_M	1,353.92	1,980.47	39.49
MixQ5A6E8-v0	1,336.72	1,963.61	37.43
Q6_K	1,251.84	1,816.35	34.56

CUDA prompt-processing medians exceed Vulkan by +13% to +20% for all three artifacts; tg128 medians differ by +1–3%. The candidate's Vulkan-only pp2048 advantage over Q5_K_M (+2.34%) did not reproduce under CUDA (−0.85% here), while its margin over Q6_K widened. The scoped classification above is unchanged: a measured intermediate, nondominated point — on either backend. Evidence: results/llama-bench-frontier-q4toq8-cuda-b10442-20260819T094256Z and results/llama-bench-mixq5a6e8-v0-cuda-b10442-20260818T154432Z (an earlier 3-model CUDA session; values reproduced within ~1–2%).

Measured Q4–Q8 frontier (CUDA lane)

Six artifacts, identical sealed protocol (CUDA b10442, dual RTX 3090, layer split 1/1, pp512/pp2048/tg128, 5 repetitions, medians). BF16-relative KL and full WikiText-2 PPL are from the pinned Vulkan perplexity lane (identical protocol across artifacts; Q8_0 was not KL/PPL-measured here).

Artifact	Bytes	BF16-relative KL ↓	Full PPL ↓	`pp512` tok/s ↑	`pp2048` tok/s ↑	`tg128` tok/s ↑
Q4_K_M	17,772,537,440	0.009584	6.6921	1,406.90	2,103.70	42.47
MixQ4A6E8-v0 (sibling candidate)	19,021,793,888	0.007835	6.6830	1,368.37	2,011.01	40.49
Q5_K_M	19,231,099,520	0.003949	6.6715	1,353.92	1,980.47	39.49
MixQ5A6E8-v0 (this artifact)	20,804,373,120	0.003083	6.6674	1,336.72	1,963.61	37.43
Q6_K	22,082,529,920	0.002099	6.6553	1,251.84	1,816.35	34.56
Q8_0	29,116,388,960	not measured	not measured	1,515.64	2,186.57	28.58

BF16 reference PPL is 6.648031. Within this sealed six-artifact set, the ladder is a monotone quality↔throughput trade-off: no artifact is simultaneously higher-fidelity (KL) and faster (tg128) than another, so every measured point is nondominated. MixQ5A6E8-v0 occupies the measured intermediate position between Q5_K_M and Q6_K. Caveat: Q6_K and Q8_0 each had one outlier pp2048 repetition (stddev 91/134 tok/s); medians are reported throughout.

This does not establish a broad optimization, best, recommendation, VRAM benefit, Q8/BF16 equivalence, or industry-benchmark quality claim. Prompt processing is compute-bound on this host (Q8_0 leads pp* despite its size); token generation is bandwidth-bound (ordering follows artifact size).

The fixed structural and text integration smoke suite passed; it is not an industry quality benchmark. Evidence: results/text-validation-mixq5a6e8-v0-vulkan-b10442-20260816T0810Z.

MTP: separate, unbundled, experimental

MTP is disabled by default. The Q4_0 and Q8_0 MTP draft models are separate upstream companion files and are not included in this upload. In the sealed 8K test both MTP modes first diverged from target-only at generated token 117; exact token-ID parity passed at 32K. Therefore MTP is not lossless or token-identical and must not be used for exact, schema-constrained, regression, or reproducibility-sensitive work. Evidence: results/mtp-parity-v021-mixq5a6e8-v0-vulkan-b10442-20260816T1045Z.

Multimodal: failed experimental gate

The matching Q8_0 projector is a separate upstream companion and is not included in this upload. In sealed run results/multimodal-v02-mixq5a6e8-v0-vulkan-b10442-20260816T1145Z, the candidate and Q5_K_M reference both passed projector CLI load, server startup, telemetry capture, object/color/count, chart/table, and the text-only regression with the projector loaded. Both returned LEFT_OF and ABOVE for the spatial fixture but failed its exact verbose-label scorer, and both read LIMA-47 as LINA-47. The full projector integration gate therefore failed. These matching fixture outcomes show no candidate-specific regression on this narrow suite; they do not validate general multimodal quality or support any quality or optimization claim.

Compatibility and limitations

Evidence was collected on a dedicated dual-GPU validation host with 2× NVIDIA RTX 3090 24 GiB GPUs, pinned llama.cpp b10442 Vulkan tools, and a pinned b10442 CUDA sm_86 build; other runtimes and hardware are unverified.
The tensor-selection rules remain experimental even though this exact artifact was built and experimentally checked.
Published companion artifacts, if desired, must be fetched separately from their pinned upstream revision and verified by hash.
Review MANIFEST.json, NOTICE, and ATTRIBUTION.md before redistribution.

Sibling candidate

MixQ4A6E8-v0 (Q4_K_M fallback, same attention/SSM and embedding rules) is built and measured on the same evidence base; see Alogotron/Qwen3.8-27B-MixQ4A6E8-GGUF (link activates when published).